[ci] feat: add npu unit test

[yyyy2000]

What does this PR do?

Add npu unit test workflow. Enable unit tests to auto-detect device (GPU/NPU)

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)

[gemini-code-assist]

Code Review

This pull request introduces NPU support for unit tests by abstracting device-specific logic into utility functions. The changes primarily involve replacing hardcoded "cuda" references with calls to these new utilities like get_device_name. While the overall approach is correct, I've identified a few critical issues in the implementation of some new helper functions and their usage. These could lead to test failures in environments without NPU support. I've provided detailed comments and code suggestions to address these problems and improve the robustness of the device detection logic.

[gemini-code-assist]

This new helper function get_device_name_ray has two critical issues that will cause tests to crash:

1. Unsafe NPU check: The call torch.npu.is_available() will raise an AttributeError if the torch.npu module is not present (e.g., in a CUDA-only environment).
2. UnboundLocalError: If neither CUDA nor NPU is available, the device variable is never assigned, leading to an UnboundLocalError on the return statement.

Please use a safe method to check for NPU availability and ensure there's a fallback return value (e.g., "CPU") to prevent crashes.

Suggested change

	def get_device_name_ray() -> str:
	"""Function that gets the torch.device based on the current machine.
	This currently only supports CPU, CUDA, NPU.
	Returns:
	device
	"""
	if torch.cuda.is_available():
	device = "GPU"
	elif torch.npu.is_available():
	device = "NPU"
	return device
	def get_device_name_ray() -> str:
	"""Function that gets the torch.device based on the current machine.
	This currently only supports CPU, CUDA, NPU.
	Returns:
	device
	"""
	from verl.utils.device import is_npu_available
	if torch.cuda.is_available():
	return "GPU"
	elif is_npu_available:
	return "NPU"
	return "CPU"

[yyyy2000]

已修改

[gemini-code-assist]

The call torch.npu.is_available() is unsafe and will raise an AttributeError if the torch.npu module is not available (e.g., in a CUDA-only or CPU-only environment). This will cause the test to crash. Please use a safe way to check for NPU availability, for example by using get_device_name from verl.utils.device, to ensure the code runs correctly across different hardware environments.

Suggested change

	if torch.cuda.is_available():
	torch.cuda.set_device(cls.rank)
	cls.device = torch.device(f"cuda:{cls.rank}")
	elif torch.npu.is_available():
	torch.npu.set_device(cls.rank)
	cls.device = torch.device(f"npu:{cls.rank}")
	else:
	cls.device = torch.device("cpu")
	from verl.utils.device import get_device_name
	device_name = get_device_name()
	if device_name == "cuda":
	torch.cuda.set_device(cls.rank)
	cls.device = torch.device(f"cuda:{cls.rank}")
	elif device_name == "npu":
	torch.npu.set_device(cls.rank)
	cls.device = torch.device(f"npu:{cls.rank}")
	else:
	cls.device = torch.device("cpu")

[yyyy2000]

已修改

[gemini-code-assist]

For consistency and clarity, it's better to explicitly specify the device with its index, similar to how it's done in the _worker_mask function in this same file (device=f"{get_device_name()}:{rank}"). While the current code works because set_device is called earlier, being explicit avoids reliance on implicit state and makes the code easier to understand and maintain.

Suggested change

	local = torch.tensor([float(rank + 1)], device=get_device_name())
	local = torch.tensor([float(rank + 1)], device=f"{get_device_name()}:{rank}")

[yyyy2000]

已修改